Notes about training GAN

  • Try more complex generator
  • ...so simpler the discriminator
  • A "stronger" dataset: the style of the distribution should be more highlighted

G' and G''

Exercises 7.41 and 8.14 from Rubinstein's book Polymer Physics. From Boltzmann superposition principle, we have σ(t)=tG(tt)γ˙(t)dt γ(t)=γ0exp(Iωt) σ(t)=Iωγ0tG(tt)exp(Iωt)dt=Iωγ0exp(Iωt)0G(τ)exp(Iωτ)dτ:=γ(t)IωG^(ω) We therefore have IωG^=IωFcos{G}+ωFsin{G}IG+G Fcos and Fsin represent Fourier cosine and sine transform. Take Fcos(1)=δ(ω) and Fsin(1)=1/ω (note the normalization factor of F), let G(t)=Geq+G1(t): G=ωFsin(G1+Geq)=Geq+ωFsin(G1(t))G=ωFcos(G1+Geq)=ωFcos(G1(t))+ωδ(ω)=ωFcos(G1(t)) G and G are defined as the Fourier sine and cosine transform of stress relaxation function G(t): G:=ω0G(t)sinωtdtG:=ω0G(t)cosωtdt Let G(t) be G(t)(tτ0)12exp(tτR) G' i.e., the cosine transform is thus G/ω=0dt(tτ0)12exp(tτR)cosωt Since τR=N2τ0, we have (let t2τR=t) G/ω=2τRN0dtexp(t2)cosωτRt2=2τRNRe0dtexp(t2)exp(IωτRt2)=2τRNRe0dtexp((1IωτR)t2) Since t2=(t)2, extending the range of last integral to (,) yields a Gaussian integral with complex argument, the result is straightforward:
G/ω=τRNReπ1IωτR=τRNReπ1+IωτR1+(ωτR)2=τRNReπ11+(ωτR)21+(ωτR)2×(cos(12arctanωτR)+Isin(12arctanωτR))=τRNπ11+(ωτR)21+(ωτR)2cos(12arctanωτR) With cos(arctan(x))=11+x2 and 2cos(x/2)21=cos(x), we finally have G=ωτRπ21+(ωτR)2+11+(ωτR)2 The factor π2 comes from the definition of Fourier cosine transform.
P.S. It is more convinient (non-restrictively) to treat the integral tαexp(t)exp(Iwt) with 0<α<1 as Gamma function with (complex) coefficient in exp: 0dttαexp((1Iω)t)=(1Iωt)α1Γ(1α) We may use Hypergeometric function pFq for more genernal cases, i.e., tαexp(tβ)exp(Iwt)
Back2Top ^